[cekees]

C2: retire XDMF/XML from proteus

Only ymf2xmf needs XDMF now. The archive's writers build ymf grid dicts directly; nothing in the proteus package imports an XML library.

Needs cekees/ymf#20 merged first (CI installs ymf from @main): the two XDMF reads a version 1 archive still needs moved there.

What changed

• The running archive is dicts, not a tree. Archiver.ArchiveDomain holds named ArchiveCollections, each with only the current step's ymf grid dict (the tree used to be emptied at every sync too). write_grid starts a step's grid and records its time on the collection; write_topology, write_geometry and write_field fill it with ymf's own builders and return the data item. allGatherIncremental stores the dicts as they are, so the parse_grid_element bridge is gone, and the stored step time is the float, not a %e string.
• Every writer converted: the 14 writeMeshXdmf_* and the hand-built function writers in Archiver.XdmfWriter; Mesh, Mesh2DM and Mesh3DM.writeMeshXdmf in MeshTools (the base mesh, its element-boundary mesh, the local-to-global maps and the material types); AuxiliaryVariables.writeScalarXdmf (the value is inlined in the document). Dataset names and sidecar names are unchanged, so existing archives read as before. The "does this time already have a grid?" test every writer made against its grid's <Time> is now collection.is_current(t).
• Gone: self.tree, indentXML, from xml.etree.ElementTree import * in Archiver, AuxiliaryVariables and MeshTools, and the two lines in NumericalSolution that created the <Domain>.
• The mesh readers work on the domain dict: MeshTools.readMeshFromArchive reads a proteus archive from its .h5 metadata through readArchiveDomain (version 1 and 2 alike, no .xmf needed), and readMeshXdmf reads XDMF written by other tools (the OpenFOAM hex_cube_3x3 fixture) through ymf.xdmf.read_xdmf; findXMLgridElement became findArchiveGrid. The v1 fragments and the v1 collection order are parsed by ymf.xdmf.read_grid_fragment and ymf.xdmf.collection_names.
• Tests: test_archiver_write_field.py pins the dicts instead of the XML (and the builders); test_archive_no_xml.py checks by AST that no module on the archive path imports xml and that no SubElement is left in the package; the read-path, dtype, parallel-metadata and bathymetry tests lost their tree fixtures.

Fixed along the way

Four writers only ever worked on one rank under global_sync and now contribute the owned slice with offsets like the rest: writeVectorFunctionXdmf_MonomialDGPK, writeVectorFunctionXdmf_LowestOrderMixed (wrote the whole local array under the global name with a non-collective create), writeVectorFunctionXdmf_P1Bubble (zero components sized by the global count against owned slices), writeFunctionXdmf_P1Bubble (offsets read off the subdomain mesh). writeMeshXdmf_DGP0 declared a mixed topology's size from self.xdmf_topology, which does not exist on the writer; readMixedElementTopologyFromXdmf had the same self. in a module function. writeMeshXdmf_MonomialDGPK and _LowestOrderMixed now set self.mesh, which their vector writers read.

Left as it was

mbd/CouplingFSI.pyx still writes its own per-body .xmf (version 1 style fragments in its own .h5) and keeps a local copy of indentXML; converting it to the archive is a follow-up. scripts/waves2xmf and velocity2xmf are Python 2 and were already dead; scripts/gatherArchives.py composes .xmf files and stays an XDMF tool.

🤖 Generated with Claude Code